Research map / Video / Action recognition
MotionSqueeze: Neural Motion Feature Learning for Video Understanding
Heeseung Kwon, Manjin Kim, Suha Kwak, Minsu Cho
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MotionSqueeze module | TSM ResNet-50 (8 frames) · Something-Something V1 | top-1 accuracy: 45.6 -> 50.9 | Improved |
| add MotionSqueeze module | TSM ResNet-50 (8 frames) · Something-Something V2 | top-1 accuracy: 58.8 -> 63.0 | Improved |
| add MotionSqueeze module | TSM ResNet-50 (8 frames) · HMDB51 | top-1 accuracy: 71.9 -> 75.8 | Improved |
| add MotionSqueeze module | TSM ResNet-50 (8 frames) · Kinetics | top-1 accuracy: 73.5 -> 75.0 | Improved |
| use kernel-soft-argmax instead of soft-argmax (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 43.8 -> 44.6 | Improved |
| add confidence map to displacement (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 44.6 -> 45.5 | Improved |
| add backward displacement tensor (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 45.5 -> 46.0 | Improved |
| insert module at res3 instead of res4 (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 42.6 -> 45.5 | Improved |
| insert module at res3 instead of res5 (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 41.1 -> 45.5 | Improved |
| insert multiple MS modules (res2,3,4) (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 45.5 -> 45.7 | No clear effect |
| fuse by addition instead of MS features only (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 38.8 -> 45.5 | Improved |
| fuse by addition instead of multiplication (ablation) | MS module · Something-Something V1 (ResNet-18, 8 frames) | top-1 accuracy: 44.5 -> 45.5 | Improved |
Closest papers by meaning
- Optical Flow Guided Feature: A Fast and Robust Motion Representation for Video Action Recognition (CVPR 2018 · similarity 0.89)
- Hierarchical Motion Understanding via Motion Programs (CVPR 2021 · similarity 0.88)
- Learning Video Representations by Transforming Time (ECCV 2020 · similarity 0.87)
- SpeedNet: Learning the Speediness in Videos (CVPR 2020 · similarity 0.87)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.87)
- DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
